__main__.py 17.5 KB
Newer Older
helloyongyang's avatar
helloyongyang committed
1
2
3
4
5
6
7
import argparse
import torch
import torch.distributed as dist
import os
import time
import gc
import json
helloyongyang's avatar
helloyongyang committed
8
import torchvision
helloyongyang's avatar
helloyongyang committed
9
10
import torchvision.transforms.functional as TF
import numpy as np
11
from contextlib import contextmanager
helloyongyang's avatar
helloyongyang committed
12
13
14
15
from PIL import Image
from lightx2v.text2v.models.text_encoders.hf.llama.model import TextEncoderHFLlamaModel
from lightx2v.text2v.models.text_encoders.hf.clip.model import TextEncoderHFClipModel
from lightx2v.text2v.models.text_encoders.hf.t5.model import T5EncoderModel
helloyongyang's avatar
helloyongyang committed
16
from lightx2v.text2v.models.text_encoders.hf.llava.model import TextEncoderHFLlavaModel
helloyongyang's avatar
helloyongyang committed
17
18

from lightx2v.text2v.models.schedulers.hunyuan.scheduler import HunyuanScheduler
19
from lightx2v.text2v.models.schedulers.hunyuan.feature_caching.scheduler import HunyuanSchedulerTaylorCaching, HunyuanSchedulerTeaCaching
helloyongyang's avatar
helloyongyang committed
20
from lightx2v.text2v.models.schedulers.wan.scheduler import WanScheduler
21
from lightx2v.text2v.models.schedulers.wan.feature_caching.scheduler import WanSchedulerTeaCaching
helloyongyang's avatar
helloyongyang committed
22
23
24

from lightx2v.text2v.models.networks.hunyuan.model import HunyuanModel
from lightx2v.text2v.models.networks.wan.model import WanModel
lijiaqi2's avatar
lijiaqi2 committed
25
26
from lightx2v.text2v.models.networks.wan.lora_adapter import WanLoraWrapper

helloyongyang's avatar
helloyongyang committed
27
28
29
30
31
32
33
from lightx2v.text2v.models.video_encoders.hf.autoencoder_kl_causal_3d.model import VideoEncoderKLCausal3DModel
from lightx2v.text2v.models.video_encoders.hf.wan.vae import WanVAE
from lightx2v.utils.utils import save_videos_grid, seed_all, cache_video
from lightx2v.common.ops import *
from lightx2v.image2v.models.wan.model import CLIPModel


lijiaqi2's avatar
lijiaqi2 committed
34
35
@contextmanager
def time_duration(label: str = ""):
36
    torch.cuda.synchronize()
lijiaqi2's avatar
lijiaqi2 committed
37
38
    start_time = time.time()
    yield
39
    torch.cuda.synchronize()
lijiaqi2's avatar
lijiaqi2 committed
40
41
42
43
    end_time = time.time()
    print(f"==> {label} start:{time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(start_time))} cost {end_time - start_time:.2f} seconds")


helloyongyang's avatar
helloyongyang committed
44
def load_models(args, model_config):
Xinchi Huang's avatar
Xinchi Huang committed
45
    if model_config["parallel_attn_type"]:
helloyongyang's avatar
helloyongyang committed
46
47
48
49
50
51
52
53
54
        cur_rank = dist.get_rank()  # 获取当前进程的 rank
        torch.cuda.set_device(cur_rank)  # 设置当前进程的 CUDA 设备
    image_encoder = None
    if args.cpu_offload:
        init_device = torch.device("cpu")
    else:
        init_device = torch.device("cuda")

    if args.model_cls == "hunyuan":
helloyongyang's avatar
helloyongyang committed
55
56
57
58
        if args.task == "t2v":
            text_encoder_1 = TextEncoderHFLlamaModel(os.path.join(args.model_path, "text_encoder"), init_device)
        else:
            text_encoder_1 = TextEncoderHFLlavaModel(os.path.join(args.model_path, "text_encoder_i2v"), init_device)
helloyongyang's avatar
helloyongyang committed
59
60
        text_encoder_2 = TextEncoderHFClipModel(os.path.join(args.model_path, "text_encoder_2"), init_device)
        text_encoders = [text_encoder_1, text_encoder_2]
helloyongyang's avatar
helloyongyang committed
61
62
        model = HunyuanModel(args.model_path, model_config, init_device, args)
        vae_model = VideoEncoderKLCausal3DModel(args.model_path, dtype=torch.float16, device=init_device, args=args)
helloyongyang's avatar
helloyongyang committed
63
64

    elif args.model_cls == "wan2.1":
65
66
67
68
69
70
71
72
73
74
        with time_duration("Load Text Encoder"):
            text_encoder = T5EncoderModel(
                text_len=model_config["text_len"],
                dtype=torch.bfloat16,
                device=init_device,
                checkpoint_path=os.path.join(args.model_path, "models_t5_umt5-xxl-enc-bf16.pth"),
                tokenizer_path=os.path.join(args.model_path, "google/umt5-xxl"),
                shard_fn=None,
            )
            text_encoders = [text_encoder]
lijiaqi2's avatar
lijiaqi2 committed
75
76
77
78
79
80
81
82
83
84
85
86
        with time_duration("Load Wan Model"):
            model = WanModel(args.model_path, model_config, init_device)

        if args.lora_path:
            lora_wrapper = WanLoraWrapper(model)
            with time_duration("Load LoRA Model"):
                lora_name = lora_wrapper.load_lora(args.lora_path)
                lora_wrapper.apply_lora(lora_name, args.strength_model)
                print(f"Loaded LoRA: {lora_name}")

        with time_duration("Load WAN VAE Model"):
            vae_model = WanVAE(vae_pth=os.path.join(args.model_path, "Wan2.1_VAE.pth"), device=init_device, parallel=args.parallel_vae)
Dongz's avatar
Dongz committed
87
        if args.task == "i2v":
lijiaqi2's avatar
lijiaqi2 committed
88
89
90
91
92
93
94
            with time_duration("Load Image Encoder"):
                image_encoder = CLIPModel(
                    dtype=torch.float16,
                    device=init_device,
                    checkpoint_path=os.path.join(args.model_path, "models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth"),
                    tokenizer_path=os.path.join(args.model_path, "xlm-roberta-large"),
                )
helloyongyang's avatar
helloyongyang committed
95
96
97
98
99
100
    else:
        raise NotImplementedError(f"Unsupported model class: {args.model_cls}")

    return model, text_encoders, vae_model, image_encoder


helloyongyang's avatar
helloyongyang committed
101
def set_target_shape(args, image_encoder_output):
Dongz's avatar
Dongz committed
102
    if args.model_cls == "hunyuan":
helloyongyang's avatar
helloyongyang committed
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
        if args.task == "t2v":
            vae_scale_factor = 2 ** (4 - 1)
            args.target_shape = (
                1,
                16,
                (args.target_video_length - 1) // 4 + 1,
                int(args.target_height) // vae_scale_factor,
                int(args.target_width) // vae_scale_factor,
            )
        elif args.task == "i2v":
            vae_scale_factor = 2 ** (4 - 1)
            args.target_shape = (
                1,
                16,
                (args.target_video_length - 1) // 4 + 1,
                int(image_encoder_output["target_height"]) // vae_scale_factor,
                int(image_encoder_output["target_width"]) // vae_scale_factor,
            )
Dongz's avatar
Dongz committed
121
122
123
124
    elif args.model_cls == "wan2.1":
        if args.task == "i2v":
            args.target_shape = (16, 21, args.lat_h, args.lat_w)
        elif args.task == "t2v":
helloyongyang's avatar
helloyongyang committed
125
126
127
128
129
130
131
132
            args.target_shape = (
                16,
                (args.target_video_length - 1) // 4 + 1,
                int(args.target_height) // args.vae_stride[1],
                int(args.target_width) // args.vae_stride[2],
            )


helloyongyang's avatar
helloyongyang committed
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
def generate_crop_size_list(base_size=256, patch_size=32, max_ratio=4.0):
    num_patches = round((base_size / patch_size) ** 2)
    assert max_ratio >= 1.0
    crop_size_list = []
    wp, hp = num_patches, 1
    while wp > 0:
        if max(wp, hp) / min(wp, hp) <= max_ratio:
            crop_size_list.append((wp * patch_size, hp * patch_size))
        if (hp + 1) * wp <= num_patches:
            hp += 1
        else:
            wp -= 1
    return crop_size_list


def get_closest_ratio(height: float, width: float, ratios: list, buckets: list):
    aspect_ratio = float(height) / float(width)
    diff_ratios = ratios - aspect_ratio

    if aspect_ratio >= 1:
        indices = [(index, x) for index, x in enumerate(diff_ratios) if x <= 0]
    else:
        indices = [(index, x) for index, x in enumerate(diff_ratios) if x > 0]

    closest_ratio_id = min(indices, key=lambda pair: abs(pair[1]))[0]
    closest_size = buckets[closest_ratio_id]
    closest_ratio = ratios[closest_ratio_id]

    return closest_size, closest_ratio


helloyongyang's avatar
helloyongyang committed
164
165
def run_image_encoder(args, image_encoder, vae_model):
    if args.model_cls == "hunyuan":
helloyongyang's avatar
helloyongyang committed
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
        img = Image.open(args.image_path).convert("RGB")
        origin_size = img.size

        i2v_resolution = "720p"
        if i2v_resolution == "720p":
            bucket_hw_base_size = 960
        elif i2v_resolution == "540p":
            bucket_hw_base_size = 720
        elif i2v_resolution == "360p":
            bucket_hw_base_size = 480
        else:
            raise ValueError(f"i2v_resolution: {i2v_resolution} must be in [360p, 540p, 720p]")

        crop_size_list = generate_crop_size_list(bucket_hw_base_size, 32)
        aspect_ratios = np.array([round(float(h) / float(w), 5) for h, w in crop_size_list])
        closest_size, closest_ratio = get_closest_ratio(origin_size[1], origin_size[0], aspect_ratios, crop_size_list)

        resize_param = min(closest_size)
        center_crop_param = closest_size

        ref_image_transform = torchvision.transforms.Compose(
            [torchvision.transforms.Resize(resize_param), torchvision.transforms.CenterCrop(center_crop_param), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize([0.5], [0.5])]
        )

        semantic_image_pixel_values = [ref_image_transform(img)]
        semantic_image_pixel_values = torch.cat(semantic_image_pixel_values).unsqueeze(0).unsqueeze(2).to(torch.float16).to(torch.device("cuda"))

        img_latents = vae_model.encode(semantic_image_pixel_values, args).mode()

        scaling_factor = 0.476986
        img_latents.mul_(scaling_factor)

        target_height, target_width = closest_size

        return {"img": img, "img_latents": img_latents, "target_height": target_height, "target_width": target_width}

Dongz's avatar
Dongz committed
202
    elif args.model_cls == "wan2.1":
helloyongyang's avatar
helloyongyang committed
203
204
        img = Image.open(args.image_path).convert("RGB")
        img = TF.to_tensor(img).sub_(0.5).div_(0.5).cuda()
gushiqiao's avatar
gushiqiao committed
205
        clip_encoder_out = image_encoder.visual([img[:, None, :, :]], args).squeeze(0).to(torch.bfloat16)
helloyongyang's avatar
helloyongyang committed
206
207
208
        h, w = img.shape[1:]
        aspect_ratio = h / w
        max_area = args.target_height * args.target_width
Dongz's avatar
Dongz committed
209
210
        lat_h = round(np.sqrt(max_area * aspect_ratio) // args.vae_stride[1] // args.patch_size[1] * args.patch_size[1])
        lat_w = round(np.sqrt(max_area / aspect_ratio) // args.vae_stride[2] // args.patch_size[2] * args.patch_size[2])
helloyongyang's avatar
helloyongyang committed
211
212
213
214
215
        h = lat_h * args.vae_stride[1]
        w = lat_w * args.vae_stride[2]

        args.lat_h = lat_h
        args.lat_w = lat_w
Dongz's avatar
Dongz committed
216
217

        msk = torch.ones(1, 81, lat_h, lat_w, device=torch.device("cuda"))
helloyongyang's avatar
helloyongyang committed
218
        msk[:, 1:] = 0
Dongz's avatar
Dongz committed
219
        msk = torch.concat([torch.repeat_interleave(msk[:, 0:1], repeats=4, dim=1), msk[:, 1:]], dim=1)
helloyongyang's avatar
helloyongyang committed
220
221
        msk = msk.view(1, msk.shape[1] // 4, 4, lat_h, lat_w)
        msk = msk.transpose(1, 2)[0]
gushiqiao's avatar
gushiqiao committed
222
223
224
        vae_encode_out = vae_model.encode(
            [torch.concat([torch.nn.functional.interpolate(img[None].cpu(), size=(h, w), mode="bicubic").transpose(0, 1), torch.zeros(3, 80, h, w)], dim=1).cuda()], args
        )[0]
helloyongyang's avatar
helloyongyang committed
225
226
227
228
        vae_encode_out = torch.concat([msk, vae_encode_out]).to(torch.bfloat16)
        return {"clip_encoder_out": clip_encoder_out, "vae_encode_out": vae_encode_out}

    else:
gushiqiao's avatar
gushiqiao committed
229
        raise NotImplementedError(f"Unsupported model class: {args.model_cls}")
helloyongyang's avatar
helloyongyang committed
230
231


helloyongyang's avatar
helloyongyang committed
232
def run_text_encoder(args, text, text_encoders, model_config, image_encoder_output):
helloyongyang's avatar
helloyongyang committed
233
234
235
    text_encoder_output = {}
    if args.model_cls == "hunyuan":
        for i, encoder in enumerate(text_encoders):
helloyongyang's avatar
helloyongyang committed
236
237
238
239
            if args.task == "i2v" and i == 0:
                text_state, attention_mask = encoder.infer(text, image_encoder_output["img"], args)
            else:
                text_state, attention_mask = encoder.infer(text, args)
Dongz's avatar
Dongz committed
240
241
            text_encoder_output[f"text_encoder_{i + 1}_text_states"] = text_state.to(dtype=torch.bfloat16)
            text_encoder_output[f"text_encoder_{i + 1}_attention_mask"] = attention_mask
helloyongyang's avatar
helloyongyang committed
242
243
244
245
246
247
248
249
250
251
252
253
254
255

    elif args.model_cls == "wan2.1":
        n_prompt = model_config.get("sample_neg_prompt", "")
        context = text_encoders[0].infer([text], args)
        context_null = text_encoders[0].infer([n_prompt if n_prompt else ""], args)
        text_encoder_output["context"] = context
        text_encoder_output["context_null"] = context_null

    else:
        raise NotImplementedError(f"Unsupported model type: {args.model_cls}")

    return text_encoder_output


helloyongyang's avatar
helloyongyang committed
256
def init_scheduler(args, image_encoder_output):
helloyongyang's avatar
helloyongyang committed
257
258
    if args.model_cls == "hunyuan":
        if args.feature_caching == "NoCaching":
helloyongyang's avatar
helloyongyang committed
259
            scheduler = HunyuanScheduler(args, image_encoder_output)
260
261
        elif args.feature_caching == "Tea":
            scheduler = HunyuanSchedulerTeaCaching(args, image_encoder_output)
helloyongyang's avatar
helloyongyang committed
262
        elif args.feature_caching == "TaylorSeer":
263
            scheduler = HunyuanSchedulerTaylorCaching(args, image_encoder_output)
helloyongyang's avatar
helloyongyang committed
264
265
266
267
268
269
270
        else:
            raise NotImplementedError(f"Unsupported feature_caching type: {args.feature_caching}")

    elif args.model_cls == "wan2.1":
        if args.feature_caching == "NoCaching":
            scheduler = WanScheduler(args)
        elif args.feature_caching == "Tea":
271
            scheduler = WanSchedulerTeaCaching(args)
helloyongyang's avatar
helloyongyang committed
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
        else:
            raise NotImplementedError(f"Unsupported feature_caching type: {args.feature_caching}")

    else:
        raise NotImplementedError(f"Unsupported model class: {args.model_cls}")
    return scheduler


def run_main_inference(args, model, text_encoder_output, image_encoder_output):
    for step_index in range(model.scheduler.infer_steps):
        torch.cuda.synchronize()
        time1 = time.time()

        model.scheduler.step_pre(step_index=step_index)

        torch.cuda.synchronize()
        time2 = time.time()

        model.infer(text_encoder_output, image_encoder_output, args)

        torch.cuda.synchronize()
        time3 = time.time()

        model.scheduler.step_post()

        torch.cuda.synchronize()
        time4 = time.time()

        print(f"step {step_index} infer time: {time3 - time2}")
        print(f"step {step_index} all time: {time4 - time1}")
        print("*" * 10)

    return model.scheduler.latents, model.scheduler.generator


def run_vae(latents, generator, args):
    images = vae_model.decode(latents, generator=generator, args=args)
    return images


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--model_cls", type=str, required=True, choices=["wan2.1", "hunyuan"], default="hunyuan")
    parser.add_argument("--task", type=str, choices=["t2v", "i2v"], default="t2v")
    parser.add_argument("--model_path", type=str, required=True)
    parser.add_argument("--config_path", type=str, default=None)
    parser.add_argument("--image_path", type=str, default=None)
Dongz's avatar
Dongz committed
319
    parser.add_argument("--save_video_path", type=str, default="./output_ligthx2v.mp4")
helloyongyang's avatar
helloyongyang committed
320
321
322
323
324
325
326
327
328
    parser.add_argument("--prompt", type=str, required=True)
    parser.add_argument("--infer_steps", type=int, required=True)
    parser.add_argument("--target_video_length", type=int, required=True)
    parser.add_argument("--target_width", type=int, required=True)
    parser.add_argument("--target_height", type=int, required=True)
    parser.add_argument("--attention_type", type=str, required=True)
    parser.add_argument("--sample_neg_prompt", type=str, default="")
    parser.add_argument("--sample_guide_scale", type=float, default=5.0)
    parser.add_argument("--sample_shift", type=float, default=5.0)
Dongz's avatar
Dongz committed
329
330
331
332
333
    parser.add_argument("--do_mm_calib", action="store_true")
    parser.add_argument("--cpu_offload", action="store_true")
    parser.add_argument("--feature_caching", choices=["NoCaching", "TaylorSeer", "Tea"], default="NoCaching")
    parser.add_argument("--mm_config", default=None)
    parser.add_argument("--seed", type=int, default=42)
Xinchi Huang's avatar
Xinchi Huang committed
334
    parser.add_argument("--parallel_attn_type", default=None, choices=["ulysses", "ring"])
Dongz's avatar
Dongz committed
335
336
337
338
    parser.add_argument("--parallel_vae", action="store_true")
    parser.add_argument("--max_area", action="store_true")
    parser.add_argument("--vae_stride", default=(4, 8, 8))
    parser.add_argument("--patch_size", default=(1, 2, 2))
helloyongyang's avatar
helloyongyang committed
339
340
    parser.add_argument("--teacache_thresh", type=float, default=0.26)
    parser.add_argument("--use_ret_steps", action="store_true", default=False)
lijiaqi2's avatar
lijiaqi2 committed
341
342
343
    parser.add_argument("--use_bfloat16", action="store_true", default=True)
    parser.add_argument("--lora_path", type=str, default=None)
    parser.add_argument("--strength_model", type=float, default=1.0)
helloyongyang's avatar
helloyongyang committed
344
345
346
347
    args = parser.parse_args()

    start_time = time.time()
    print(f"args: {args}")
Dongz's avatar
Dongz committed
348

helloyongyang's avatar
helloyongyang committed
349
350
    seed_all(args.seed)

Xinchi Huang's avatar
Xinchi Huang committed
351
    if args.parallel_attn_type:
Dongz's avatar
Dongz committed
352
        dist.init_process_group(backend="nccl")
helloyongyang's avatar
helloyongyang committed
353
354
355
356
357
358
359

    if args.mm_config:
        mm_config = json.loads(args.mm_config)
    else:
        mm_config = None

    model_config = {
360
        "model_cls": args.model_cls,
helloyongyang's avatar
helloyongyang committed
361
362
363
364
365
366
367
        "task": args.task,
        "attention_type": args.attention_type,
        "sample_neg_prompt": args.sample_neg_prompt,
        "mm_config": mm_config,
        "do_mm_calib": args.do_mm_calib,
        "cpu_offload": args.cpu_offload,
        "feature_caching": args.feature_caching,
Xinchi Huang's avatar
Xinchi Huang committed
368
        "parallel_attn_type": args.parallel_attn_type,
Dongz's avatar
Dongz committed
369
        "parallel_vae": args.parallel_vae,
lijiaqi2's avatar
lijiaqi2 committed
370
        "use_bfloat16": args.use_bfloat16,
helloyongyang's avatar
helloyongyang committed
371
372
373
374
375
376
377
378
379
    }

    if args.config_path is not None:
        with open(args.config_path, "r") as f:
            config = json.load(f)
        model_config.update(config)

    print(f"model_config: {model_config}")

lijiaqi2's avatar
lijiaqi2 committed
380
381
    with time_duration("Load models"):
        model, text_encoders, vae_model, image_encoder = load_models(args, model_config)
382

Dongz's avatar
Dongz committed
383
    if args.task in ["i2v"]:
helloyongyang's avatar
helloyongyang committed
384
385
386
387
        image_encoder_output = run_image_encoder(args, image_encoder, vae_model)
    else:
        image_encoder_output = {"clip_encoder_out": None, "vae_encode_out": None}

388
389
    with time_duration("Run Text Encoder"):
        text_encoder_output = run_text_encoder(args, args.prompt, text_encoders, model_config, image_encoder_output)
helloyongyang's avatar
helloyongyang committed
390

helloyongyang's avatar
helloyongyang committed
391
392
    set_target_shape(args, image_encoder_output)
    scheduler = init_scheduler(args, image_encoder_output)
helloyongyang's avatar
helloyongyang committed
393
394
395
396
397
398
399

    model.set_scheduler(scheduler)

    gc.collect()
    torch.cuda.empty_cache()
    latents, generator = run_main_inference(args, model, text_encoder_output, image_encoder_output)

400
401
402
403
    if args.cpu_offload:
        scheduler.clear()
        del text_encoder_output, image_encoder_output, model, text_encoders, scheduler
        torch.cuda.empty_cache()
helloyongyang's avatar
helloyongyang committed
404

405
406
    with time_duration("Run VAE"):
        images = run_vae(latents, generator, args)
helloyongyang's avatar
helloyongyang committed
407

Xinchi Huang's avatar
Xinchi Huang committed
408
    if not args.parallel_attn_type or (args.parallel_attn_type and dist.get_rank() == 0):
409
410
411
412
413
        with time_duration("Save video"):
            if args.model_cls == "wan2.1":
                cache_video(tensor=images, save_file=args.save_video_path, fps=16, nrow=1, normalize=True, value_range=(-1, 1))
            else:
                save_videos_grid(images, args.save_video_path, fps=24)
helloyongyang's avatar
helloyongyang committed
414
415

    end_time = time.time()
416
    print(f"Total cost: {end_time - start_time}")